AI does not get caught. Its mistakes do.

Most of the argument about AI in IB coursework is about detection. That is the wrong argument, and it is the reason students keep losing marks they did not expect to lose.

Start by conceding the detection point

Text detectors are unreliable in both directions. They clear work that was generated and they flag work that was not, and a student who writes in a plain, orderly register can be accused on no evidence at all. Nobody should plan their coursework around a tool that behaves like that, in either direction.

So set detection aside. It is not what will cost you marks.

What will cost you marks

Language models produce fluent text and, now and then, confident falsehoods. A citation formatted perfectly for a paper that does not exist. A constant that is nearly right. A study summarised as saying the opposite of what it says. A mechanism described in the right vocabulary and the wrong order.

These survive your own reading for a specific reason. You cannot notice an error in the exact area where you were outsourcing your understanding. If you had the knowledge to catch it, you would have written the sentence yourself. The errors land precisely in the places you are least equipped to audit.

Then two readers see it, and they are not the same reader

Your teacher reads your work knowing you. They saw the practical. They have heard you explain the concept correctly in class. When a sentence is ambiguous they resolve it charitably, because they know what you meant, and that charity is invisible to you, because it never appears as a comment.

The moderator or examiner has none of that. They have the document. When your bibliography lists a paper, they are the person who might look for it. When your discussion rests on a figure, they are the person who knows what that figure should be. Moderation compares your school's marking against the IB standard, and a piece of work that reads well but does not hold up is exactly where the two markers diverge.

Under the Extended Essay criteria, an argument built on a wrong premise loses twice: once on knowledge and understanding, and again on discussion and evaluation, because you cannot evaluate evidence you have got wrong. The fluency does not rescue it. Fluency is not a criterion.

And the misconduct question, separately

The IB has not banned AI tools. It requires that their use is acknowledged. Copy, paraphrase or modify anything an AI produced and it is cited, in the body of the work and in the bibliography, in your school's referencing style, with the prompt you gave and the date. Work produced even in part by such a tool is not regarded as your own. Hiding the use is academic misconduct, with the consequences that carries.

Both things are true at once. The misconduct rule is the one with the dramatic penalty; the accuracy problem is the one that quietly takes marks off far more students, in far more essays, without anyone ever opening an investigation.

The usable rule

If you cannot independently verify a claim, it does not go in your essay, whatever produced it. That rule is older than these tools, it applies equally to a forum post and a textbook you half remember, and it is the whole of the defence.

For what we do and do not use AI for in your feedback, our policy is here: the short version is that no AI writes your report, and no human edits your text.

Sources

IB position on AI tools, the acknowledgement requirement and the prompt-and-date citation format: the IB's published statement on AI in assessment and education, and the IB academic integrity policy. Checked 25 August 2026. Extended Essay criteria are described in our own words and not reproduced. See your subject guide. Note that the EE criteria changed for first assessment in May 2027; November 2026 candidates sit the previous version.